Prediction of atmospheric motion vectors around typhoons using generative adversarial network

Abstract

In this study, atmospheric motion vectors (AMVs) were derived from the satellite images predicted using a generative adversarial network (GAN) and a deep multi-scale frame prediction algorithm. The GAN was trained and tested with a sequence of the satellite images of a COMS satellite infrared-window channel under the 68 tropical cyclones. The inputs of the consecutive satellite images with 15-min interval were then processed using the trained GAN model to generate satellite images in the next time steps. To further enhance the model's predictability, particle image velocimetry based on the theory of cross-correlation schemes was employed to the GAN-generated satellite image sequence and AMVs were produced. The GAN-derived AMVs were validated with the wind fields based on the numerical weather prediction (NWP) and radiosonde observations. The comparisons showed that the GAN-derived AMVs depicted the structure of atmospheric circulations with a certain level of accuracy. Through comparison with the radiosonde observations, the root-mean-square error and the wind speed bias of the GAN-derived AMVs were comparable to, and even smaller than those of the NWP-derived wind fields. The current approach may enhance the accuracy in predicting short-term wind velocity fields, which in turn may provide more realistic inputs in storm surge modeling.